FMAFS: camouflaged object detection via fusion of multiscale adaptive-frequency and spatial domain features

Camouflaged object detection (COD) is a challenging task in the field of computer vision. When dealing with camouflaged objects that are highly similar to complex backgrounds, the detection performance of existing methods is limited, and they are susceptible to interference from illumination variations and background noise. To overcome these limitations, we explored the application of frequency-domain features in camouflaged object detection and proposed a multi-scale adaptive frequency-domain and spatial-domain feature fusion model (FMAFS) based on learnable frequency-domain transformations. We designed three adaptive frequency-domain encoders: Learnable Fast Fourier Transform (LFFT), Learnable Discrete Cosine Transform (LDCT), and Learnable Discrete Wavelet Transform (LDWT). These encoders could adaptively adjust the frequency-domain feature extraction approach based on the data, enhancing the flexibility and accuracy of feature extraction. By integrating these learnable frequency-domain transformations with spatial-domain features, we constructed a model with an encoder-decoder architecture that effectively captured subtle features of camouflaged objects, significantly improving detection performance. Experimental results demonstrated that the adaptive frequency-domain encoder based on LFFT achieved the best performance on three available public datasets (CAMO, COD10K, NC4K), significantly outperforming other existing COD models.